TL;DR
PREFER (PRompt Ensemble learning via Feedback-REflect-Refine) is a new framework that treats prompts as "weak learners" in an automated boosting ensemble. By using an LLM to self-reflect on its own errors and iteratively refine its instructions, PREFER achieves state-of-the-art performance across reasoning and classification tasks without manual prompt tuning.
What Problem Does This Paper Address?
Prompt engineering has become a cornerstone of working with Large Language Models (LLMs), but it remains a "black art"—highly manual, trial-and-error prone, and sensitive to minor wording changes. While "prompt ensembling" (using multiple prompts and voting on the result) helps reduce instability and hallucinations, existing methods are often static. They require a pre-defined set of prompts and don't allow the model to learn from its specific mistakes.
The authors identify a major gap: existing ensemble methods don't perform "directed optimization." They don't look at which examples the model is failing on and specifically update the prompts to handle those "hard" cases.
Methods
PREFER introduces a two-pronged approach to make prompt ensembling automatic and adaptive:
- Feedback-Reflect-Refine (The Boosting Branch): Inspired by the classical AdaBoost algorithm, PREFER treats each prompt as a "weak learner."
- Feedback: The system identifies "hard examples" (instances the current prompt got wrong).
- Reflect: The LLM is given these errors and asked to explain why the current prompt failed.
- Refine: Based on this reflection, the LLM generates a new, more nuanced prompt designed to solve those specific hard cases. This creates a "textual gradient" that guides the optimization process.
- Bilateral Bagging (The Stability Branch): To make individual prompts more reliable, the authors propose "bilateral bagging." Instead of just asking for an answer, the model performs:
- Forward Thinking: Evaluating the confidence that a candidate answer is correct.
- Backward Thinking: Evaluating the confidence that a candidate answer should be excluded (process of elimination). By combining these scores, the system counteracts the known tendency of LLMs to be overconfident in their predictions.
Datasets
The researchers evaluated PREFER on seven diverse benchmarks:
- Natural Language Inference (NLI): SNLI, MNLI, QNLI, and RTE (tasks requiring the model to determine if one sentence entails another).
- Classification: Ethos (hate speech detection), Liar (fake news classification), and ArSarcasm (Arabic sarcasm detection).
Key Insights and Ideas
- Text as a Gradient: One of the most compelling ideas is treating natural language feedback as a "gradient." In traditional neural networks, we use numerical gradients to update weights. Here, the LLM’s self-reflection serves as the directional signal to "update" the prompt.
- Targeting the "Hard" Space: By focusing on examples that previous iterations failed on, the ensemble naturally expands its "solving space." Each new prompt acts as a specialist for a different sub-problem.
- Calibration via Elimination: The "Bilateral Bagging" insight mirrors human decision-making. When we are unsure, we often look for reasons to rule things out. Implementing this "backward thinking" significantly improves the model's calibration and reduces variance.
Key Results
PREFER consistently outperformed single-prompt baselines, Chain-of-Thought (CoT) prompting, and existing state-of-the-art ensemble methods like PromptBoosting and APO.
- Accuracy Gains: On the Liar dataset (fake news), PREFER improved performance by 13.1% over the next best method. On QNLI, it saw a 6.3% boost.
- Efficiency: Compared to APO (Automatic Prompt Optimization), PREFER reached peak performance in fewer iterations and required significantly fewer API calls (see Table 3 in the paper), making it more cost-effective for production environments.
- Ablation Success: The authors showed that removing the "Feedback" component led to a severe performance drop (over 10% in many cases), proving that directed self-reflection is the "secret sauce" of the framework.
Limitations and Future Work
While PREFER is more efficient than some iterative methods, it still requires multiple LLM calls per inference to run the ensemble, which increases latency and cost compared to a single prompt.
The authors suggest two main directions for future work:
- Cost Reduction: Finding ways to distill the knowledge of the ensemble back into a single, high-performing "super-prompt."
- Algorithm Evolution: Exploring how other classical machine learning algorithms (beyond boosting and bagging) can be "re-imagined" using the natural language reasoning capabilities of LLMs.
Summary updated